GReS: Graphical Cross-domain Recommendation for Supply Chain Platform

Zhiwen Jing, Ziliang Zhao, Yan Feng, Xaochen Ma, Nan Wu, Shengqiao Kang, Cheng Yang, Yujia Zhang, Hao Guo

Abstract: Supply Chain Platforms (SCPs) provide downstream industries with numerous raw materials. Compared with traditional e-commerce platforms, data in SCPs is more sparse due to limited user interests. To tackle the data sparsity problem, one can apply Cross-Domain Recommendation (CDR) which improves the recommendation performance of the target domain with the source domain information. However, applying CDR to SCPs directly ignores the hierarchical structure of commodities in SCPs, which reduce the recommendation performance. To leverage this feature, in this paper, we take the catering platform as an example and propose GReS, a graphical cross-domain recommendation model. The model first constructs a tree-shaped graph to represent the hierarchy of different nodes of dishes and ingredients, and then applies our proposed Tree2vec method combining GCN and BERT models to embed the graph for recommendations. Experimental results on a commercial dataset show that GReS significantly outperforms state-of-the-art methods in Cross-Domain Recommendation for Supply Chain Platforms.

TL;DR

Supply Chain Platforms (SCPs) struggle with severe data sparsity, making accurate product recommendations difficult. This paper introduces GReS, a novel Cross-Domain Recommendation (CDR) model that leverages data from downstream consumer platforms by mapping the hierarchical relationships between finished products (like restaurant dishes) and raw materials (like ingredients). By combining Graph Convolutional Networks (GCNs) and a customized BERT model, GReS significantly improves recommendation accuracy for SCPs.

What Problem Does This Paper Address?

Supply Chain Platforms (SCPs) provide raw materials to downstream industries, such as selling wholesale ingredients to restaurants. Recommender systems are crucial for these platforms, but they suffer from extreme "data sparsity." Because business users have very narrow, concentrated purchasing behaviors compared to standard retail consumers, there isn't enough interaction data to generate high-quality recommendations.

A common fix for data sparsity is Cross-Domain Recommendation (CDR), which borrows user data from a richer, related domain (like a consumer-facing food delivery app) to improve recommendations in the sparse domain (the SCP). However, existing CDR methods treat items across domains as equals. They fail to capture the hierarchical relationship inherent in supply chains. For instance, a "Beef Burger" sold in the downstream domain is actually composed of "beef," "tomatoes," and "pickles" bought in the SCP domain. Ignoring this compositional hierarchy leads to sub-optimal recommendations.

Methods

To bridge the gap between the SCP (Domain A) and the downstream catering platform (Domain B), the authors propose GReS (Graphical Cross-domain Recommendation for Supply Chain Platform). The model architecture is organized into five distinct layers to process and combine cross-domain information.

The five-layer architecture of GReS, integrating heterogeneous and tree-shaped graphs for cross-domain recommendation.
  1. Graph Structure Layer: Domain A (SCP) is modeled as a Heterogeneous Graph (HG) containing users, items, and categories. Domain B (Catering) is modeled as a Tree-shaped Graph (TG) built from Tree-shaped Meta Structures (TMSs) to capture the Dish $\rightarrow$ Category $\rightarrow$ Ingredient hierarchy.
  2. Graph Embedding Layer: HG nodes are embedded using Node2vec. For the TG, the authors propose a novel Tree2vec method that fuses Graph Convolutional Networks (GCNs) for spatial features and BERT for semantic features. To adapt BERT for trees, they use Depth-First Search (DFS) to flatten the tree and modify position embeddings to represent hierarchical levels.
  3. Feature Combination Layer: An element-wise attention layer merges embeddings from both domains, dynamically weighting source domain information based on the sparsity of the user's data in the target domain.
  4. Neural Network Layer: Deep Structured Semantic Models (DSSM) process the combined features through multiple ReLU layers.
  5. Prediction Layer: The final layer generates the recommendation score $\hat{y}_{ih}$.

Key Results

Because there are no public datasets for SCP CDR, the authors used a proprietary dataset from a commercial online system containing 3,000 common users and 600 SCP-only users.

Performance comparison of GReS against baselines and ablation study results.
  • Overall Performance: GReS outperformed five baselines, including state-of-the-art cross-domain models like DCDIR. It achieved an average improvement of 9.03% across metrics such as HR@K, NDCG@K, and MRR@K.
  • Ablation Studies:
    • Removing the tree structure caused a 17.36% performance drop, highlighting that hierarchy is a game-changer for supply chain recommendations.
  • Removing GCN or BERT components resulted in drops of 9.57% and 9.68%, respectively, proving that both structural and semantic learning are essential.
  • Removing custom fixed-position embeddings in BERT caused a 1.80% drop, validating the need to explicitly encode tree depth.

Key Insights and Future Work

  • Hierarchy Matters: Treating cross-domain items as a compositional hierarchy rather than a flat 1:1 mapping is vital for SCPs.
  • Customizing NLP for Graphs: By using DFS traversal and tweaking position embeddings, the authors successfully adapted a transformer model to understand hierarchical graph semantics.
  • Future Directions: The authors suggest extending GReS beyond the catering industry to other SCP environments with similar hierarchical structures, such as industrial accessories, medical care, and commodity wholesale.

Summary updated

Abstract

Supply Chain Platforms (SCPs) provide downstream industries with raw materials. Compared with traditional e-commerce platforms, data in SCPs is more sparse due to limited user interests. To tackle the data sparsity problem, one can apply Cross-Domain Recommendation (CDR) to improve the recommendation performance of the target domain with the source domain information. However, applying CDR to SCPs directly ignores hierarchical structures of commodities in SCPs, which reduce recommendation performance. In this paper, we take the catering platform as an example and propose GReS, a graphical CDR model. The model first constructs a tree-shaped graph to represent the hierarchy of different nodes of dishes and ingredients, and then applies our proposed Tree2vec method combining GCN and BERT models to embed the graph for recommendations. Experimental results show that GReS significantly outperforms state-of-the-art methods in CDR for SCPs.
BibTeX
@inproceedings{jing2022gres,
  author = {Zhiwen Jing and Ziliang Zhao and Yan Feng and Xaochen Ma and Nan Wu and Shengqiao Kang and Cheng Yang and Yujia Zhang and Hao Guo},
  title = {{GReS: Graphical Cross-domain Recommendation for Supply Chain Platform}},
  booktitle = {Proceedings of the 31st ACM International Conference on Information \& Knowledge Management},
  year = {2022},
  doi = {10.1145/3511808.3557607},
}
Figure